Compressed CNN for Atrial Fibrillation Detection on Wearable Devices

Monitoring patients' electrocardiograms (ECG) with smart wearable devices for real-time detection of atrial fibrillation (AF) can significantly improve healthcare outcomes. Convolutional neural networks (CNNs) are highly effective in detecting such anomalous heartbeats, yet their high computational complexity poses challenges for deployment on resource-constrained platforms. In this work, we present an efficient approach to deploying lightweight CNN-based AF detection models on wearable devices. We developed an advanced classification net-work based on MobileNetV2, trained and validated on the ECG dataset from the Intelligent Embedded System Design Contest at ICMC. To improve model generalization, we applied GAN-based data augmentation techniques. Additionally, to optimize the model's size, we employed a combined CNN compression approach that incorporates Dependency Graph Pruning and Quantization-Aware Training. Inference experiments conducted on the Loongson 2K0500 development board highlighted the superior performance of our approach. The proposed method enabled real-time inference with a low latency of 9.67 ms and a compact model size of 41.1 KB, while maintaining a high accuracy of 96.52 %. These results demonstrate that our CNN compression scheme is exceptionally well-suited for deploying AF detection models on resource-constraint wearable devices.